Abstract
Traditional Zenith Tropospheric Delay (ZTD) models often face difficulties in maintaining prediction accuracy under complex meteorological conditions and data loss. To address this, we propose the transformer-xLSTM (TransXLT) model, which integrates spatial-temporal information from global navigation satellite system (GNSS) stations, ERA5 (global atmospheric reanalysis), and GPT3 (empirical ZTD estimation). Missing data are reconstructed using a sparse attention-based time series reconstruction (SASR) method. Experimental results show: (1) under a 120-h data loss, SASR reduces mean absolute error (MAE) by 24.5% compared to cubic Hermite interpolation; (2) SASR lowers training root mean square error (RMSE) by 15.1% versus direct data deletion; and (3) TransXLT achieves an average RMSE of 8.13 mm across six sites, reducing RMSE by up to 76.54% compared to benchmarks like CNN-LSTM and ERA5. Demonstrating robustness across varying latitudes, altitudes, and seasons, the model significantly advances ZTD estimation accuracy for GNSS applications.